In-situ PV generation forecasting of utility-scale solar power plants in Ghana: Single and hybrid machine learning approaches

Abstract

• Solar irradiation, ambient temperature and PV module temperature are key factors and significant variables to estimate PV power generation. • ANN, RF and XGBoost algorithms are accurate power prediction models for large scale PV plants in Ghana. • XGBoost time series forecasting for PV plants is reliable for 1-day ahead and 1-month ahead prediction. • XGBoost-ANN hybrid model forecasting for PV plants is reliable for 1-week and 2-weeks ahead prediction. • Hybrid and ensemble models are more accurate than single models, to implement as intelligent solar forecasting technology in smart-grid systems. The intrinsic intermittency and weather-dependent fluctuations make renewable energy (RE) integration into the electrical grid challenging. Thus, short-to-medium term forecasting of power generation of utility-scale photovoltaic (PV) systems is key for reliable operational planning for efficient grid-power management. The study evaluated and compared six machine learning models to predict PV generation based on three positive Pearson correlated input weather features: solar irradiation, PV module temperature and ambient temperature. The evaluated models comprise eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Bayesian Regression (BR), Support Vector Regression (SVR), Linear Regression (LR), and Artificial Neural Network (ANN). Subsequently, hybrid models composed of RF-ANN, RF-XGBoost and XGBoost-ANN stacked from ANN, RF and XGBoost as the best performing models among all the trained predictive models were built to forecast PV generation: one day, one week, two weeks and one month ahead. The models were validated on data from a 50 MWp PV system at Bui, Ghana. Employing performance evaluation metrics including the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE). The study findings revealed, for both day-ahead and month-ahead forecasts, XGBoost outperformed all the other models. XGBoost-ANN dominated the performance for the week-ahead and two-weeks-ahead forecasts. Comparative with existing works, these findings present practical insights and reliable results with normalized errors deemed acceptable for operational purposes in the scope of RE forecasting. Reinforcing policy frameworks for stakeholders aiming at adopting intelligent solar forecasting technology in smart-grid systems, to enhance operational planning and energy management for grid resilience and reliability.